After your AutoML tabular classification model is done training, create an endpoint and deploy your model to the endpoint. After your model is deployed to this new endpoint, test your model by requesting a prediction.
Load your model
When your model finishes training, it is listed in the Models tab.
In the Google Cloud console, in the Vertex AI section, go to the Models page.
From the models list, click the name of your trained model that you created previously
Models are organized into versions. Click model version number 1.
Evaluate your model
The Evaluate panel helps you understand how the model performed against the test set. When you are done, continue to the next part of the tutorial.
Optional. Hold the pointer over the
? icons to learn about each evaluation
Optional. Move the confidence threshold slider to see how the precision, recall, and F1 scores are affected.
The confusion matrix shows how a prediction compares to the test set (ground truth).
Recall that label "1" is the negative class (the customer did not sign up for a term deposit) and "2" is the positive class. Your model likely did a better job predicting the negative class than the positive class. Perhaps with additional training time, more data, or additional features, you could improve predictive performance for the positive class.
Feature importance shows how each feature impacted model training: The higher the value, the more impactful.
Your model probably shows that duration (how long the most recent communication between the bank and customer lasted, in seconds) contributed heavily to the prediction outcome.
Deploy your model to an endpoint
To test a model or make online predictions, you need to deploy it to an endpoint.
Open the Deploy & Test panel.
Under Deploy your model, click Deploy to endpoint.
Structured_AutoML_Tutorialfor the endpoint name.
Keep the minimum compute node at
1and don't enter a maximum.
Turn off model monitoring for this endpoint.
Click Deploy. to create your endpoint and deploy your model to it.
Model deployment takes around 5 minutes. When your endpoint is ready, proceed to the next part of the tutorial.
Request a prediction
Now that your model is deployed to an endpoint, you can send prediction requests. Rather than send a request through the API or gcloud, you can test your model on this page.
In the Test your model section, you'll see a Value column that's pre-filled. You can use those values or enter new ones.
At the bottom of the section, press Predict.
For this model, a prediction result of
1represents a negative outcome—a deposit is not made at the bank. A prediction result of
2represents a positive outcome—a deposit is made at the bank.
Your model will return a confidence score, which is the model's level of certainty that the selected label is the correct one. The default value probably returned a high confidence score.
Optional. Try changing duration to a much higher value and press Predict again.
Follow the last page of the tutorial to clean up resources that you have created.
Learn more about model evaluation.
Learn more about model predictions.